11  Probability Applied — Sensor Noise from First Principles

The previous chapter laid out the generic measurement chain. This page works it through end-to-end for one concrete sensor — the camera — because every link corresponds to physically distinct hardware, making it the cleanest worked example. The same template applies to vibration, audio, network telemetry; the table at the bottom shows where each one plugs in.


11.1 1. Random variables and distributions — recap

A random variable is a function that maps an experiment’s outcome to a number. Every sensor reading is a random variable — it takes a different value each time you measure the same underlying state, because the physics of measurement is inherently random.

Three quantities characterise a distribution:

  • Mean \mu = E[X] — the expected value
  • Variance \sigma^2 = E[(X - \mu)^2] — the spread
  • Standard deviation \sigma — same units as X

11.2 2. The distribution chain — worked for the camera

11.2.1 Bernoulli — one photon, one photosite

Each photon either reaches the photosite or it doesn’t. Probability p of detection (= quantum efficiency, QE):

P(X = 1) = p, \qquad P(X = 0) = 1 - p

Mean = p, variance = p(1 - p).

Generic version: one event, two outcomes, probability p of success.

11.2.2 Binomial — n photons, one photosite

Count the number detected in n independent trials:

P(X = k) \;=\; \binom{n}{k} p^k (1 - p)^{n - k}

Mean = np, variance = np(1 - p).

Generic version: successes in n identical independent trials.

11.2.3 Poisson — photon flux model

When n \to \infty and p \to 0 such that np = \lambda stays fixed:

P(X = k) \;=\; \frac{\lambda^k \, e^{-\lambda}}{k!}

Mean = \lambda, variance = \lambda. This is the exact model for photon arrivals over an exposure.

Generic version: rare events arriving at rate \lambda per window — photons per exposure, shock pulses per second, packet drops per minute.

11.2.4 Normal — large-count limit

When \lambda \gg 1, the Poisson distribution converges to a Gaussian:

\text{Poisson}(\lambda) \;\xrightarrow{\lambda \to \infty}\; \mathcal{N}(\mu = \lambda, \sigma^2 = \lambda)

This is why shot noise is Gaussian in bright conditions.

11.2.5 Central Limit Theorem — why sums are always Gaussian

In a sensor, multiple noise sources add together (shot + read + dark + quantization). Even if each source is non-Gaussian, the CLT says the sum is approximately Gaussian. This is why total sensor noise is well-modelled as Gaussian in almost all practical cases — and the same conclusion holds for a vibration sensor, an audio ADC, or any other measurement front-end.


11.3 3. Shot noise is irreducible

From the Poisson model: \sigma_{\text{shot}} = \sqrt\lambda.

This cannot be reduced by better hardware — it is a fundamental property of the discrete nature of light. To halve the shot noise you must quadruple the photon count (larger aperture, longer exposure, bigger photosite).

\text{SNR} \;=\; \frac{\lambda}{\sqrt\lambda} \;=\; \sqrt\lambda

Doubling the photon count improves SNR by \sqrt 2 \approx 41\,\%. The same \sigma = \sqrt\lambda rule holds for any Poisson process: bearing-fault counts, packet drops, Geiger-counter clicks.


11.4 4. Noise budget

A real sensor has multiple noise sources. The total noise variance is their sum (assuming independence):

\sigma_{\text{total}}^2 \;=\; \sigma_{\text{shot}}^2 + \sigma_{\text{read}}^2 + \sigma_{\text{dark}}^2 + \sigma_{\text{quant}}^2 \;=\; \lambda + \sigma_r^2 + \sigma_d^2 + \frac{\Delta^2}{12}

where \Delta is the quantization step size.

Which term dominates depends on the signal regime:

Regime Dominant noise Characteristic
Bright (\lambda \gg \sigma_r^2) Shot noise \sigma \propto \sqrt\lambda
Dim (\lambda \ll \sigma_r^2) Read noise \sigma \approx \sigma_r (constant floor)
Saturated (\lambda \geq C) Clipping All information above C is lost

In bright conditions, collecting more signal always helps. In dim conditions, electronics dominate and more signal doesn’t help much until you climb out of the electronics floor.

The same regime structure shows up in any measurement chain: an electronics-limited floor at low signal, a physics-limited (often \sqrt\lambda) regime at medium signal, and a saturation regime where the measurement is destroyed.


11.5 5. The Anscombe transform

Poisson noise has variance equal to its mean — awkward statistically. The Anscombe transform converts Poisson noise to approximately unit-variance Gaussian:

A(X) \;=\; 2\sqrt{X + \tfrac{3}{8}}

After applying A, the data has approximately constant variance \approx 1 regardless of \lambda. This lets standard Gaussian-assumption algorithms work correctly on Poisson-distributed data — useful any time you have count data you want to feed into a tool that assumes constant-variance Gaussian noise.


11.6 6. The same chain — where other sensors plug in

Stage Camera Vibration (accelerometer) Network telemetry Audio (microphone)
Bernoulli event Photon detected (QE) Shock pulse in a time slice Packet dropped or not Pressure crossing threshold
Binomial count Photons detected per exposure Pulses per window Drops per minute Threshold crossings per frame
Poisson regime \lambda = expected photons \lambda = expected pulses/hour \lambda = expected drops/min (rarely the dominant model)
Gaussian limit High illumination Many pulses per window High-traffic monitoring Almost always — the noise floor is Gaussian to begin with
Electronics floor Read noise Charge-amp + ADC noise (none) Pre-amp + ADC noise
Quantization ADC bit depth ADC bit depth (count, no quantization) 16/24-bit ADC

The chain is the same. Only the names change.


11.7 Summary

Concept Key fact
Bernoulli One trial, two outcomes
Binomial n trials, count of successes
Poisson n \to \infty, p \to 0, np = \lambda; mean = variance = \lambda
Gaussian Limiting case \lambda \to \infty — why most noise floors are approximately Normal
CLT Sum of any independent RVs → Gaussian; why total sensor noise is Normal
SNR = \sqrt\lambda in any Poisson regime (photons, pulses, packets)
Noise budget \sigma_{\text{total}}^2 = \sigma_{\text{physics}}^2 + \sigma_{\text{electronics}}^2 + \sigma_{\text{quant}}^2